See axolotl config

axolotl version: 0.6.0

base_model: PrimeIntellect/INTELLECT-1-Instruct
trust_remote_code: true
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

gpu_memory_limit: 

load_in_8bit: 
load_in_4bit:
strict: false

chat_template: llama3
datasets:
  - path: neginashz/rationale-llama-chat-dataset
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content

    
dataset_prepared_path:
val_set_size: 0.1
output_dir: ./star-sft-intellect-2




sequence_len: 8192
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true


wandb_project: star-sft-intellect-instruct-2
wandb_entity: 
wandb_watch:
wandb_name: 
wandb_log_model: 

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 2
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00002

train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false

gradient_checkpointing: true
  
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps:
eval_steps: 
save_steps:

evals_per_epoch: 16
saves_per_epoch: 1

debug:
deepspeed: deepspeed_configs/zero2.json
weight_decay:
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|end_of_text|>
  pad_token: <|finetune_right_pad_id|>

hub_model_id: neginashz/star-sft-intellect-instruct-2
hub_strategy: 
early_stopping_patience:

resume_from_checkpoint:
auto_resume_from_checkpoints: true

star-sft-intellect-instruct-2

This model is a fine-tuned version of PrimeIntellect/INTELLECT-1-Instruct on the neginashz/rationale-llama-chat-dataset dataset. It achieves the following results on the evaluation set:

• Loss: 0.3719

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• total_train_batch_size: 4
• total_eval_batch_size: 4
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 6
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
0.5033	0.0686	7	0.4057
0.4303	0.1373	14	0.3986
0.4496	0.2059	21	0.3977
0.4223	0.2745	28	0.3973
0.4083	0.3431	35	0.3940
0.4191	0.4118	42	0.3893
0.412	0.4804	49	0.3859
0.3912	0.5490	56	0.3812
0.3995	0.6176	63	0.3749
0.4236	0.6863	70	0.3703
0.3833	0.7549	77	0.3663
0.3605	0.8235	84	0.3614
0.3952	0.8922	91	0.3576
0.3744	0.9608	98	0.3540
0.199	1.0196	105	0.3536
0.1762	1.0882	112	0.4128
0.1704	1.1569	119	0.3808
0.1603	1.2255	126	0.3781
0.1727	1.2941	133	0.3874
0.1624	1.3627	140	0.3841
0.1546	1.4314	147	0.3793
0.1602	1.5	154	0.3776
0.1501	1.5686	161	0.3745
0.146	1.6373	168	0.3734
0.1512	1.7059	175	0.3733
0.146	1.7745	182	0.3725
0.1479	1.8431	189	0.3721
0.1395	1.9118	196	0.3720
0.1472	1.9804	203	0.3719

Framework versions

• Transformers 4.47.0
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.21.0